Pith. sign in

REVIEW 4 major objections 5 minor 40 references

Efficient Transformer-Inspired Variants of Physics-Informed Deep Operator Networks

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

Pith's one-line read Six simple input-swap modifications close most of the gap between vanilla and modified DeepONet accuracy at a fraction of the training cost.

desk verdict Useful practical variants and careful empirical tables, but the abstract overclaims and test-set-based variant selection inflates the central 'matches or surpasses' claim. read the letter →

arxiv 2509.01679 v1 pith:HZOVTT27 submitted 2025-09-01 cs.LG cs.NAmath.NAstat.ML

classification cs.LGcs.NAmath.NAstat.ML
keywords physics-informedoperatorlearningDeepONettransformer-inspiredarchitecturescross-conditioningpartialdifferentialequationstrainingefficiencyFourierfeaturesbenchmarkcomparison
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper targets a practical trade-off in physics-informed operator learning: the simple vanilla DeepONet trains quickly but is less accurate, while the modified DeepONet is more accurate but takes two to three times longer to train. The authors propose six variants that keep the vanilla architecture and cost but let the branch network (which encodes the input function) receive the query point's spatial coordinate, and let the trunk network (which encodes the query point) receive information about the input function. On four benchmarks—advection, diffusion-reaction, Burgers', and Korteweg-de Vries—each equation has a variant whose accuracy matches or beats the modified DeepONet while training roughly 40-60% faster per iteration. The winning variant tracks the physics: full velocity field for advection, local source value for diffusion-reaction, leading Fourier coefficients for Burgers', and full initial condition for KdV. If this holds, users can get the accuracy of the expensive architecture without paying its training cost.

What carries the argument

The central mechanism is bidirectional cross-conditioning between the branch and trunk networks of DeepONet. The branch net's coefficients b(u,x) are made query-dependent by adding the spatial coordinate x to its inputs; the trunk net's basis functions gamma(t,x,·) are made context-aware by adding either the local input value u(x), the full input field u, or a truncated set of Fourier coefficients û_Λ to its inputs. This mirrors the dynamic query-context coupling of Transformer attention, but as a change of network inputs rather than an added attention module. Six variants—Bx, TL, BxTL, BxTG, TF, BxTF—span the combinations, and the paper's experiments identify which conditioning pattern fits

What would settle it

Generate a fresh test set from the same Gaussian random field distributions, pre-specify the best variant per equation without inspecting test errors, retrain all six variants and the modified DeepONet under identical budgets, and compare median relative L2 errors; if the selected variant no longer falls within the paper's TOST equivalence margins of the modified DeepONet on any benchmark, the central claim fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that Transformer-inspired bidirectional cross-conditioning narrows the accuracy gap between the vanilla and modified DeepONets. In the vanilla DeepONet the branch and trunk networks are independent; the authors inject the query coordinate x into the branch, making the coefficients depend on where the solution is evaluated, and inject input-function context into the trunk, making the basis functions depend on the problem instance. The resulting variants are non-intrusive—they only change network inputs—and remain simple enough to preserve the vanilla model's training efficiency. Measured against the modified DeepONet on four PDE benchmarks, the best variant per eq

Load-bearing premise

The reported 'matches or surpasses' record depends on the authors choosing the best variant and loss-weighting scheme for each equation after seeing results on the same test set, so the advantage could shrink under a strictly held-out selection protocol.

Editorial extensions

If this is right

  • For each benchmark, a variant reaches modified-DeepONet-level accuracy while using roughly 40-60% less wall-clock time to complete the training budget (e.g., BxTG finishes advection in 2,929 s vs 4,470 s; BxTG finishes KdV in 5,837 s vs 14,667 s).
  • For periodic problems with Fourier boundary conditions, trunk conditioning on leading Fourier coefficients—TF and BxTF—is the top choice; for non-periodic or full-field problems, full-function conditioning BxTG wins.
  • The low-viscosity Burgers' regime is the clearest accuracy win: BxTF's 12.1% mean relative L2 error beats the modified DeepONet's 18.1% while training at about 60% of the per-iteration cost.
  • Wilcoxon TOST equivalence tests support treating several of these matches as statistically equivalent, and high Spearman correlations show the variants err on the same test cases as the baseline.
  • Since the design only changes network inputs, the variants work with existing physics-informed loss weighting schemes (CK and BRDR) without new training algorithms.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: if the match-up holds on held-out data, a practical design rule emerges—choose conditioning by PDE type (Fourier conditioning for periodic problems, full-field conditioning for hyperbolic and dispersive problems) rather than by trial-and-error.
  • The paper itself points to localized domain-of-dependence conditioning as future work; a direct test would be conditioning the trunk on u(x±ε) for hyperbolic equations and checking whether accuracy improves further or merely matches full-field conditioning.
  • The same cross-conditioning idea should transfer to other branch-trunk operator architectures, since it does not depend on DeepONet-specific training details.
  • A stronger test of the physics-alignment claim would be to fix the conditioning rule a priori for a new PDE family and check that the predicted best variant is indeed best.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper proposes six Transformer-inspired variants of the physics-informed DeepONet (Bx, TL, BxTL, BxTG, TF, BxTF) that inject query-point information into the branch network and input-function information into the trunk network. The authors evaluate these variants on four PDE benchmarks—advection, diffusion–reaction, Burgers' equation at three viscosities, and KdV—comparing accuracy and training efficiency against the 'vanilla' and modified DeepONets. They report that for each case there exists a variant that matches or surpasses the accuracy of the modified DeepONet while reducing training time, and that the best-performing variant for each equation aligns with the equation's physical characteristics. The empirical methodology includes multiple random seeds, raw data tables, Wilcoxon TOST equivalence tests, Glass's Delta effect sizes, and Spearman rank correlations.

Significance. If the central claim were established, the paper would offer a useful, simple architectural recipe for improving the speed–accuracy trade-off of physics-informed DeepONets, with the non-intrusive cross-conditioning idea being sufficiently novel to interest the operator-learning community. Strengths of the paper include careful reproducibility practices: raw numerical data are provided in Appendix B.1, training and data-generation parameters are tabulated in Appendices A and D, and the statistical protocol is detailed in Appendix C. The use of multiple seeds and nonparametric tests is commendable. However, two issues currently prevent the central claim from being accepted: one is a direct internal contradiction at Burgers' nu=1e-2, and the other is a selection-on-the-test-set bias that undermines the 'exists a variant' formulation. The empirical evidence is valuable, but the paper's headline claim is not yet supported by the reported analysis.

major comments (4)
  1. [§3.3 and Table C1] The abstract and Section 4 claim that 'for each case, there exists a variant that matches or surpasses the accuracy of the modified DeepONet.' This is directly contradicted by the Burgers' nu=1e-2 results. Table C1 reports that TOST rejects equivalence, with median difference 0.0248% favoring the modified DeepONet and Glass's Delta = 0.397. Table B3 shows that every variant has a larger mean relative L2 error than the modified DeepONet (e.g., TF 0.220% vs. modified 0.190%). The paper itself acknowledges the accuracy compromise in Section 3.3. The abstract and Section 4 must be revised to exclude this case or to state the conditional claim (e.g., 'in all but one Burgers' case').
  2. [§3 and Appendix C] The 'exists a variant' claim is evaluated by selecting, for each equation, the variant with the lowest mean test error on the same test set used for the reported statistics and TOST tests (Figures 3, 6, 9–11, 18; Tables B1–B6). No held-out validation split or multiple-comparison correction is described. Because the accuracy of the selected variant is the maximum over several candidates, it carries a positive selection bias even if all variants are truly no better than the baseline. The TOST p-values are conditional on the chosen variant and therefore do not provide unbiased evidence for the existential claim. The physical-alignment narrative in Section 4 is built on the same selected variants and inherits this bias. A held-out evaluation of the selected variants, or a report of all variants with appropriate correction, is needed before the central claim can be regarded as established.
  3. [§3, per-equation tuning choices] In addition to variant selection, the per-equation choices of loss weighting (CK vs. BRDR) and Fourier feature embedding order are made using the same test data ('we adopt whichever loss-weighting scheme yields better accuracy, and also test deterministic and random Fourier feature embeddings'). This is another degree of freedom selected on the test set. Even if the final comparison were repeated on a truly held-out split, the selection procedure that chooses the best configuration on the evaluation data would still bias the reported accuracy. The authors should specify a protocol that separates model selection from evaluation, or explicitly quantify the optimism in the reported numbers.
  4. [§4] The claim that the best-performing variant 'aligns naturally with the equation's underlying physical characteristics' is a post-hoc interpretation of the selected variants and is not supported by any predictive experiment. For example, the explanation that TF works for Burgers because low-frequency modes dominate is plausible but is not tested by, say, perturbing the Fourier truncation order and observing a monotone effect, or by comparing against a variant that injects an equivalent amount of information in a non-physical form. As written, this is a suggestive narrative rather than an evidence-backed conclusion; it should be framed more cautiously or supported by additional controlled experiments.
minor comments (5)
  1. [§3.4] Typo: 'lower perdictive error' should be 'lower predictive error.'
  2. [Appendix B.2, Figures B1–B15] Several captions refer to a 'representative training instance,' but the text and surrounding discussion indicate these are test instances. The captions should be corrected.
  3. [Appendix C.2.3] The equivalence margin Delta = 0.2 * min_i epsilon_i^(b) is data-dependent. This is a reasonable anchoring rule, but its interpretation should be clarified: the margin and the resulting TOST p-values are conditional on the observed baseline errors, and the margin is not a pre-specified constant. A sentence noting this dependence and its effect on multiple runs would help.
  4. [§3.3, Burgers' nu=1e-4 discussion] The statement that 'nonlinear energy transfer continues to favor lower modes' for viscous Burgers' is imprecise; nonlinearity tends to transfer energy to higher modes before dissipation acts. Since this claim is used only as motivation, it should be reworded or removed.
  5. [Appendix B.3] The literature comparisons report results from other papers under different training budgets, loss-balancing schemes, and initial-condition distributions. These are useful context, but they should be explicitly labeled as non-controlled comparisons so that readers do not interpret them as head-to-head benchmarks.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported accuracies are measured against external reference solutions, not derived from fitted parameters or self-cited theorems.

full rationale

The paper's central claims are empirical. For each PDE benchmark, the proposed variants are trained and their relative L2 errors are measured against reference solutions generated by independent numerical solvers (Appendix A). No equation in Section 2 defines the reported accuracy in terms of the variants' inputs or outputs, and no fitted parameter is renamed as a prediction. The equivalence margin Delta = 0.2 x min_i epsilon_i^(b) (Appendix C.2.3) is data-dependent but is a decision rule for the TOST procedure, not a fitted parameter whose value is later reported as the variant's accuracy; the variant errors themselves come from separate test-set evaluations. Self-citations, such as the BRDR weighting scheme [W. Chen et al. 2025] and the KdV setup from [Williams et al. 2024], are used as training or data-generation choices and are not load-bearing justification for the accuracy comparison; they do not import a uniqueness theorem or smuggle in the target result. The per-equation selection of the best variant and of loss-weighting/Fourier-feature settings on the same test set is a legitimate statistical concern: it introduces selection bias and means the reported 'matches or surpasses' claim is not a clean out-of-sample statement. Indeed, Table C1 shows the selected variant for Burgers (nu=1e-2) is inequivalent and worse than the modified DeepONet (Glass's Delta = 0.397), contradicting the abstract's blanket claim. That is a validity/correctness issue, but it is not circularity: the accuracy numbers are measured, not constructed, and the selected variant's error is not forced to equal or better the baseline by definition. No step in the paper's derivation chain reduces to its own inputs.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central accuracy comparisons depend on the choice of reference solvers, per-equation loss weighting and Fourier feature settings, and the statistical decision rule for equivalence. These are listed as free parameters and axioms above; none of them are derivable from the model itself. No new physical entities are introduced.

free parameters (3)
  • Equivalence margin factor = 0.2 (multiplier applied to min baseline error per equation)
    Defines Delta = 0.2 * min_i epsilon^b_i for the TOST equivalence tests in Appendix C.2.3; the resulting margins are 0.10%, 0.02%, 0.01%, 0.06%, 0.86%, 0.04% per equation. The 0.2 factor is chosen by hand and directly determines which variant-baseline pairs are declared statistically equivalent.
  • Loss-weighting scheme per equation = BRDR for advection; CK for diffusion-reaction, Burgers, KdV
    Section 3 states 'we adopt whichever loss-weighting scheme (CK or BRDR) yields better accuracy'. This is a post-hoc selection on the test performance of each equation.
  • Fourier feature embedding order per equation = random 150 draws for diffusion-reaction; deterministic orders 4, 6, 8 for Burgers nu=1e-2,1e-3,1e-4; order 12 for KdV
    Appendix A.2 lists these per-equation choices; they are hyperparameters selected to improve accuracy and are not derived from the model.
assumptions (3)
  • domain assumption The numerically generated reference solutions (Lax-Wendroff, implicit finite difference, spectral/Fourth-order ETD) are sufficiently accurate to serve as ground truth for the reported relative L2 errors.
    All error metrics in Section 3 are computed against these reference solutions; if the references were inaccurate, the comparisons would not be valid. Appendix A.1 describes the solvers.
  • domain assumption The physics-informed training with CK or BRDR loss weighting converges for all variants within the fixed iteration budget, and multiple-seed averaging adequately controls stochasticity.
    The paper reports means over random seeds but does not show a convergence proof; the comparisons assume that each model reached a representative optimum.
  • standard math The universal approximation theorem for DeepONet-type networks holds for the function classes considered, so the architectures have sufficient expressive capacity.
    Invoked implicitly by framing the branch/trunk outputs as basis functions and coefficients in Section 2.2. No formal capacity guarantee is proven for the specific variants.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Efficient Transformer-Inspired Variants of Physics-Informed Deep Operator Networks." pith.science (2026). https://pith.science/paper/HZOVTT27

@misc{pith2026250901679,
  author       = {Pith},
  title        = {Pith review of: Efficient Transformer-Inspired Variants of Physics-Informed Deep Operator Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HZOVTT27}},
  note         = {Machine review of arXiv:2509.01679}
}
read the original abstract

Operator learning has emerged as a promising tool for accelerating the solution of partial differential equations (PDEs). The Deep Operator Networks (DeepONets) represent a pioneering framework in this area: the "vanilla" DeepONet is valued for its simplicity and efficiency, while the modified DeepONet achieves higher accuracy at the cost of increased training time. In this work, we propose a series of Transformer-inspired DeepONet variants that introduce bidirectional cross-conditioning between the branch and trunk networks in DeepONet. Query-point information is injected into the branch network and input-function information into the trunk network, enabling dynamic dependencies while preserving the simplicity and efficiency of the "vanilla" DeepONet in a non-intrusive manner. Experiments on four PDE benchmarks -- advection, diffusion-reaction, Burgers', and Korteweg-de Vries equations -- show that for each case, there exists a variant that matches or surpasses the accuracy of the modified DeepONet while offering improved training efficiency. Moreover, the best-performing variant for each equation aligns naturally with the equation's underlying characteristics, suggesting that the effectiveness of cross-conditioning depends on the characteristics of the equation and its underlying physics. To ensure robustness, we validate the effectiveness of our variants through a range of rigorous statistical analyses, among them the Wilcoxon Two One-Sided Test, Glass's Delta, and Spearman's rank correlation.

Figures

Figures reproduced from arXiv: 2509.01679 by the authors.

Figure 1
Figure 1. Physics-informed “vanilla” DeepONet (From [Williams et al. 2024]). The branch net processes the input function u. Specifically, it takes a discrete representation of u, evaluated at a fixed set of sensor spatial locations {xi} m i=1, and produces a feature em￾bedding (b1, b2, . . . , bw) ⊺ ∈ R w. The trunk net takes the continuous spatiotemporal coordinates y = (t, x) as input and generates a corresponding feature e… view at source ↗
Figure 2
Figure 2. The modified DeepONet architecture (From [S. Wang, H. Wang et al. 2022]). Page 4 of 53 [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Performance comparison of model variants for the Advection equation. The horizontal axis shows the average per-iteration training time ratio relative to the modified DeepONet, averaged over multiple random seeds (lower is faster), while the vertical axis shows the mean relative L 2 error (lower is better). Marker shapes distinguish different model variants; the modified DeepONet is shown in black, and the best-perfo… view at source ↗
Figures from the paper (17 more)
Figure 4
Figure 4. Figure 4: Comparison of training efficiency for the Advection equation. The horizontal axis shows the wall-clock training time for a single representative run, and the vertical axis shows the mean relative L 2 error over all test cases (log scale). Dashed vertical lines and anno…
Figure 5
Figure 5. Figure 5: Distribution of relative L 2 errors for selected variants (gold: Variant BxTG; black: the modified DeepONet) on the Advection equation, evaluated over multiple random seeds and test instances. Horizontal bars span the interquartile range (25th to 75th percentile), and …
Figure 6
Figure 6. Figure 6: Performance comparison of model variants for the Diffusion–Reaction equation. The horizontal axis shows the average per-iteration training time ratio relative to the modified DeepONet, averaged over multiple random seeds (lower is faster), while the vertical axis shows…
Figure 7
Figure 7. Figure 7: Comparison of training efficiency for the Diffusion–Reaction equation. The horizontal axis shows the wall-clock training time for a single representative run, and the vertical axis shows the mean relative L 2 error over all test cases (log scale). Dashed vertical lines…
Figure 8
Figure 8. Figure 8: Distribution of relative L 2 errors for selected variants (gold: Variant TL; black: the modified DeepONet) on the Diffusion–Reaction equation, evaluated over multiple random seeds and test instances. Horizontal bars span the interquartile range (25th to 75th percentile…
Figure 9
Figure 9. Figure 9: Performance comparison of model variants for the Burgers’ equation (ν = 10−2 ). The horizontal axis shows the average per-iteration training time ratio relative to the modified DeepONet, averaged over multiple random seeds (lower is faster), while the vertical axis sho…
Figure 10
Figure 10. Figure 10: Performance comparison of model variants for the Burgers’ equation (ν = 10−3 ). The horizontal axis shows the average per-iteration training time ratio relative to the modified DeepONet, averaged over multiple random seeds (lower is faster), while the vertical axis sh…
Figure 11
Figure 11. Figure 11: Performance comparison of model variants for the Burgers’ equation (ν = 10−4 ). The horizontal axis shows the average per-iteration training time ratio relative to the modified DeepONet, averaged over multiple random seeds (lower is faster), while the vertical axis sh…
Figure 12
Figure 12. Figure 12: Comparison of training efficiency for the Burgers’ equation (ν = 10−2 ). The horizontal axis shows the wall-clock training time for a single representative run, and the vertical axis shows the mean relative L 2 error over all test cases (log scale). Dashed vertical li…
Figure 13
Figure 13. Figure 13: Comparison of training efficiency for the Burgers’ equation (ν = 10−3 ). The horizontal axis shows the wall-clock training time for a single representative run, and the vertical axis shows the mean relative L 2 error over all test cases (log scale). Dashed vertical li…
Figure 14
Figure 14. Figure 14: Comparison of training efficiency for the Burgers’ equation (ν = 10−4 ). The horizontal axis shows the wall-clock training time for a single representative run, and the vertical axis shows the mean relative L 2 error over all test cases (log scale). Dashed vertical li…
Figure 15
Figure 15. Figure 15: Distribution of relative L 2 errors for selected variants (gold: Variant TF; black: the modified DeepONet) on the Burgers’ equation (ν = 10−2 ), evaluated over multiple random seeds and test instances. Horizontal bars span the interquartile range (25th to 75th percent…
Figure 16
Figure 16. Figure 16: Distribution of relative L 2 errors for selected variants (gold: Variant TF; black: the modified DeepONet) on the Burgers’ equation (ν = 10−3 ), evaluated over multiple random seeds and test instances. Horizontal bars span the interquartile range (25th to 75th percent…
Figure 17
Figure 17. Figure 17: Distribution of relative L 2 errors for selected variants (gold: best-performing variants; black: the modified DeepONet) on the Burgers’ equation (ν = 10−4 ), evaluated over multiple random seeds and test instances. Horizontal bars span the interquartile range (25th t…
Figure 18
Figure 18. Figure 18: Performance comparison of model variants for the KdV equation. The horizontal axis shows the average per-iteration training time ratio relative to the modified DeepONet, averaged over multiple random seeds (lower is faster); the vertical axis shows the mean relative L…
Figure 19
Figure 19. Figure 19: Comparison of training efficiency for the KdV equation. The horizontal axis shows the wall-clock training time for a single representative run, and the vertical axis shows the mean relative L 2 error over all test cases (log scale). Dashed vertical lines and annotatio…
Figure 20
Figure 20. Figure 20: Distribution of relative L 2 errors for selected variants (gold: Variant BxTG; black: the modified DeepONet) on the KdV equation, evaluated over multiple random seeds and test instances. Horizontal bars span the interquartile range (25th to 75th percentile), and red d…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 19 canonical work pages

  1. [1]

    A mathematical guide to operator learning

    Boullé, Nicolas and Alex Townsend 2024 “A mathematical guide to operator learning”, in,Numerical analysis meets machine learning, ed. by Siddhartha Mishra and Alex Townsend, Handbook of Numerical Analysis, Elsevier, vol. 25, chap. 3, pp. 83–125, d o i: 10.1016/bs.hna.2024.05.003. Page 48 of 53

  2. [2]

    Positional knowledge is all you need: position-induced transformer (PiT) for operator learning

    Chen, Junfeng and Kailiang Wu 2024 “Positional knowledge is all you need: position-induced transformer (PiT) for operator learning”, in Proceedings of the 41st International Conference on Machine Learning, ICML’24, JMLR.org, pp. 7526–7552, u r l: https://dl.acm.org/doi/abs/10.5555/3692070.3692363

  3. [3]

    Self-adaptive weights based on balanced residual decay rate for physics-informed neural networks and deep operator networks

    Chen, Wenqian, Amanda A. Howard and Panos Stinis 2025 “Self-adaptive weights based on balanced residual decay rate for physics-informed neural networks and deep operator networks”,Journal of Computational Physics, article 114226, d o i: 10.1016/j.jcp.2025.114226

  4. [4]

    Exponential time differencing for stiff systems

    Cox, Stephen M. and Paul C. Matthews 2002 “Exponential time differencing for stiff systems”,Journal of Computational Physics, vol. 176, issue 2, pp. 430–455, d o i: 10.1006/jcph.2002.6995. E, Weinan and Bing Yu 2018 “The deep Ritz method: a deep learning-based numerical algorithm for solving variational problems”,Communications in Mathematics and Statisti...

  5. [5]

    19, u r l: https://bookstore.ams.org/gsm-19-r

    Evans, Lawrence C 2022 Partial Differential Equations, 2nd ed., Graduate Studies in Mathematics, Amer- ican mathematical society, vol. 19, u r l: https://bookstore.ams.org/gsm-19-r

  6. [6]

    Burgulence

    Frisch, U. and J. Bec 2001 “Burgulence”, in, New trends in turbulence turbulence: nouveaux aspects, ed. by M. Lesieur, A. Yaglom and F. David, Springer Berlin Heidelberg, Berlin, Heidelberg, pp. 341–383, d o i: 10.1007/3-540-45674-0_7

  7. [7]

    Primary, secondary, and meta-analysis of research

    Glass, Gene V. 1976 “Primary, secondary, and meta-analysis of research”,Educational Researcher, vol. 5, issue 10, pp. 3–8, u r l: https://www.jstor.org/stable/1174772

  8. [8]

    Solving high-dimensional partial differential equations using deep learning

    Han, Jiequn, Arnulf Jentzen and Weinan E 2018 “Solving high-dimensional partial differential equations using deep learning”, Proceedings of the National Academy of Sciences, vol. 115, issue 34, pp. 8505– 8510, d o i: 10.1073/pnas.1718942115. Page 49 of 53

Show all 40 references
  1. [9]

    GNOT: a general neural operator transformer for operator learning

    Hao, Zhongkai, Zhengyi Wang, Hang Su, Chengyang Ying, Yinpeng Dong, Songming Liu, Ze Cheng, Jian Song and Jun Zhu 2023 “GNOT: a general neural operator transformer for operator learning”, inPro- ceedings of the 40th International Conference on Machine Learning, ICML’23, JMLR.o...

  2. [10]

    590, d o i: 10.1007/978-3-319-22470-1

    Hesthaven, Jan S, Gianluigi Rozza and Benjamin Stamm 2016 Certified Reduced Basis Methods for Parametrized Partial Differential Equations, Springer Briefs in Mathematics, Springer, vol. 590, d o i: 10.1007/978-3-319-22470-1

  3. [11]

    Wolfe and Eric Chicken 2013 Nonparametric Statistical Methods, John Wiley & Sons, d o i: 10.1002/9781119196037

    Hollander, Myles, Douglas A. Wolfe and Eric Chicken 2013 Nonparametric Statistical Methods, John Wiley & Sons, d o i: 10.1002/9781119196037

  4. [12]

    Stacked networks improve physics-informed training: applications to neural networks and deep operator networks

    Howard, Amanda A., Sarah H. Murphy, Shady E. Ahmed and Panos Stinis 2025 “Stacked networks improve physics-informed training: applications to neural networks and deep operator networks”,Foundations of Data Science, vol. 7, issue 1, pp. 134–162, d o i: 10.3934/fods.2024029

  5. [13]

    Iserles, Arieh 2008 A First Course in the Numerical Analysis of Differential Equations, 2nd ed., Cambridge Texts in Applied Mathematics, Cambridge University Press, d o i: 10.1017/CBO9780511995569

  6. [14]

    Learning operators with coupled attention

    Kissas, Georgios, Jacob H Seidman, Leonardo Ferreira Guilhoto, Victor M Preciado, George J Pappas and Paris Perdikaris 2022 “Learning operators with coupled attention”, Journal of Machine Learning Research, vol. 23, issue 215, pp. 1–63, u r l: https://www.jmlr.org/papers/v23/2...

  7. [15]

    Neural operator: learning maps between function spaces with applications to PDEs

    Kovachki, Nikola, Zongyi Li, Burigede Liu, Kamyar Azizzadenesheli, Kaushik Bhattacharya, Andrew Stuart and Anima Anandkumar 2023 “Neural operator: learning maps between function spaces with applications to PDEs”, Journal of Machine Learning Research, vol. 24, issue 89, pp. 1–9...

  8. [16]

    Characterizing possible failure modes in physics-informed neural networks

    Krishnapriyan, Aditi, Amir Gholami, Shandian Zhe, Robert Kirby and Michael W Mahoney 2021 “Characterizing possible failure modes in physics-informed neural networks”, in Proceedings of the 35th International Conference on Neural Information Processing Systems, NIPS’21, Curran ...

  9. [17]

    Li, Zijie, Kazem Meidani and Amir Barati Farimani 2022 Transformer for partial differential equations’ operator learning, a r x i v: 2205.13671

  10. [18]

    Li, Zongyi, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart and Anima Anandkumar 2020 Fourier neural operator for parametric partial differential equations, a r x i v: 2010.08895

  11. [19]

    Learning nonlinear operators via DeepONet based on the universal approx- imation theorem of operators

    Lu, Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang and George Em Karniadakis 2021 “Learning nonlinear operators via DeepONet based on the universal approx- imation theorem of operators”,Nature Machine Intelligence, vol. 3, issue 3, pp. 218–229, d o i: 10.1038/s42256-021-00302-5

  12. [20]

    Machine-learning-based spectral methods for partial differential equations

    Meuris, Brek, Saad Qadeer and Panos Stinis 2023 “Machine-learning-based spectral methods for partial differential equations”, Scientific Reports, vol. 13, issue 1, article 1739, d o i: 10.1038/s41598-022-26602-3

  13. [21]

    Phuong, Mary and Marcus Hutter 2022 Formal algorithms for transformers, a r x i v: 2207.09238

  14. [22]

    Qadeer, Saad, Andrew Engel, Amanda Howard, Adam Tsou, Max Vargas, Panos Stinis and Tony Chiang 2023 Efficient kernel surrogates for neural network-based regression, a r x i v: 2310.18612

  15. [23]

    Quarteroni, Alfio, Andrea Manzoni and Federico Negri 2015 Reduced Basis Methods for Partial Differential Equations: An Introduction,

  16. [24]

    92, d o i: 10.1007/978-3-319-15431-2

    Unitext, Springer, vol. 92, d o i: 10.1007/978-3-319-15431-2

  17. [25]

    Physics-informed neural networks: a deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations

    Raissi, Maziar, Paris G. Perdikaris and George Em Karniadakis 2019 “Physics-informed neural networks: a deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations”,Journal of Computational Physics, vol. 378, pp. 686–707,...

  18. [26]

    A comparison of the two one-sided tests procedure and the power approach for assessing the equivalence of average bioavailability

    Schuirmann, Donald J. 1987 “A comparison of the two one-sided tests procedure and the power approach for assessing the equivalence of average bioavailability”,Journal of Pharmacokinetics and Pharmacodynamics, vol. 15, issue 6, pp. 657–680, d o i: 10.1007/BF01068419. Page 51 of 53

  19. [27]

    Transformers as neural operators for solutions of differential equations with finite regularity

    Shih, Benjamin, Ahmad Peyvan, Zhongqiang Zhang and George Em Karniadakis 2025 “Transformers as neural operators for solutions of differential equations with finite regularity”,Computer Methods in Applied Mechanics and Engineering, vol. 434, article 117560, d o i: 10.1016/j.cma...

  20. [28]

    The proof and measurement of association between two things

    Spearman, Charles 1904 “The proof and measurement of association between two things”,The American Journal of Psychology, vol. 15, issue 1, pp. 72–101, d o i: 10.2307/1412159

  21. [29]

    Tan, Lesley and Liang Chen 2022 Enhanced DeepONets for modeling partial differential operators considering multiple input functions, a r x i v: 2202.08942

  22. [30]

    Attention is all you need

    Vaswani, Ashish, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Łukasz Kaiser and Illia Polosukhin 2017 “Attention is all you need”, inProceedings of the 31st International Conference on Neural Information Processing Systems, NIPS’17, Curran Associate...

  23. [31]

    SVD perspectives for augmenting DeepONet flexibility and interpretability

    Venturi, Simone and Tiernan Casey 2023 “SVD perspectives for augmenting DeepONet flexibility and interpretability”, Computer Methods in Applied Mechanics and Engineering, vol. 403, article 115718, d o i: 10.1016/j.cma.2022.115718

  24. [32]

    Long-time integration of parametric evolution equations with physics-informed DeepONets

    Wang, Sifan and Paris Perdikaris 2023 “Long-time integration of parametric evolution equations with physics-informed DeepONets”, Journal of Computational Physics, vol. 475, article 111855, d o i: 10.1016/j.jcp.2022.111855

  25. [33]

    Learning the solution operator of parametric partial differential equations with physics-informed DeepONets

    Wang, Sifan, Hanwen Wang and Paris Perdikaris 2021 “Learning the solution operator of parametric partial differential equations with physics-informed DeepONets”,Science Advances, vol. 7, issue 40, article eabi8605, d o i: 10.1126/sciadv.abi8605. 2022 “Improved architectures an...

  26. [34]

    When and why PINNs fail to train: a neural tangent kernel perspective

    Wang, Sifan, Xinling Yu and Paris Perdikaris 2022 “When and why PINNs fail to train: a neural tangent kernel perspective”,Journal of Computational Physics, vol. 449, article 110768, d o i: 10.1016/j.jcp.2021.110768

  27. [35]

    Wellek, Stefan 2010 Testing Statistical Hypotheses of Equivalence and Noninferiority, 2nd ed., Chap- man and Hall/CRC, New York, d o i: 10.1201/EBK1439808184

  28. [36]

    2011 Introduction to Robust Estimation and Hypothesis Testing, Academic press, d o i: 10.1016/C2010-0-67044-1

    Wilcox, Rand R. 2011 Introduction to Robust Estimation and Hypothesis Testing, Academic press, d o i: 10.1016/C2010-0-67044-1

  29. [37]

    Individual comparisons by ranking methods

    Wilcoxon, Frank 1945 “Individual comparisons by ranking methods”,Biometrics Bulletin, vol. 1, issue 6, pp. 80–83, d o i: 10.2307/3001968

  30. [38]

    Williams, Emily, Amanda Howard, Brek Meuris and Panos Stinis 2024 What do physics-informed DeepONets learn? Understanding and improving training for scientific computing applications, a r x i v: 2411.18459

  31. [39]

    Yu, Xinling, Sean Hooten, Ziyue Liu, Yequan Zhao, Marco Fiorentino, Thomas Van Vaeren- bergh and Zheng Zhang 2024 Separable operator networks, a r x i v: 2407.11253

  32. [40]

    Page 53 of 53

    Zhu, Yameng, Jingrun Chen and Weibing Deng 2024 R-adaptive DeepONets: learning solution operators for PDEs with discontinuous solutions using an R-adaptive strategy, a r x i v: 2408.04157. Page 53 of 53

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.